Training on User Interruptions Builds AI Agent Robustness to Recency Bias

you can just train on "user interrupts the agent mid-task, asks for an unexpected new problem to be solved before returning to the original" for learning robustness against recency bias btw

this secretly builds pressure towards a lot of invariances at once, like domain randomization in robotics - context rot exists in a ton of task A + task B combos, and if you train on interrupts at random points on tasks of different lengths, macro deferred attention skills emerge